Zhongqing Robotics CEO: 60% of resources to brain R&D to avoid becoming 'second Unitree'

Both Zhongqing and Unitree started with legged robots, gaining attention through performances—Unitree became famous for dancing, while Zhongqing used combat to attract attention. But Zhao Tongyang only gives his flagship product, the T800, a score of 60, believing that even if the body achieves 90 or 100 points, others will still say it 'looks like Unitree.'

Unitree closed up 460% on its first day, with a P/E ratio of 219 times, but by August 28 it closed at 585 yuan per share, down 46% from its intraday high of 1100 yuan and 30% from its first-day close of 845 yuan. Zhao Tongyang realized that the 'copy Unitree' story is no longer attractive, and Zhongqing's shipments are an order of magnitude behind Unitree—Unitree's audited revenue for 2025 was 1.699 billion yuan, with 5,215 humanoid units sold.

A VP at a leading robotics company predicts the embodied intelligence industry will enter a knockout stage, where those who build large models without hardware or those with hardware but poor brains will fail. Sources close to Zhongqing previously believed that Zhongqing 'lacked brain capabilities and had no top talent,' a point Zhao Tongyang also acknowledged.

Zhao Tongyang went dark for four months, during which Zhongqing underwent a 'major resource shift': this year, 60% of funds and personnel are allocated to the brain, whereas in 2024, about 80% of manpower was on body control, with only 20% on embodied intelligence. Zhao Tongyang stated that the cerebellum's manpower is nearly sufficient, but there are still 40 points of homework to complete, which can be refined along the way, while a missing brain would lead to elimination.

In April, Zhongqing brought in Li Liyun, the 'number one' in XPeng Motors' autonomous driving division, as CTO, responsible for brain R&D. Zhao Tongyang called Li the best in China for autonomous driving, hoping to boost Zhongqing's brain from 60 to 80 points.

Boosting the brain is also a common challenge for 'Unitree-like' companies. Unitree admitted in its prospectus that its self-developed embodied large model has not yet been applied at scale, planning to invest 2.022 billion yuan to catch up. An early-stage AI investor said that companies with strong bodies are no longer attractive; the most valuable part of embodied intelligence will ultimately be the brain.

At WRC, Zhongqing released EngineAI Awaken, a five-layer brain-inspired architecture, including the instinctive reflex S1 layer, action planning S2 layer, cognitive reasoning S3 layer, emotion large model S4 layer, and self-growth S5 layer. The key design decouples 2Hz heavy semantic reasoning from 100Hz high-frequency motion control, avoiding large model latency slowing down physical responses. After integrating WAM and VLA models, with 2-hour real-machine deployment and fine-tuning, long-horizon task success rates exceed 98%.

This technical approach is similar to that of Zhipingfang and Zhuji Dynamics: the robot brain is difficult to handle with a single model; high-level models understand the environment, low-level models ensure real-time, and different layers run at different frequencies. But Zhongqing starts from a strong body, and Awaken's task is to mesh the added brain with existing joints, dexterous hands, and motion control capabilities.

Unlike Unitree, which focuses on making models more accurate, Zhongqing insists on edge inference during operation—the T800's full-body VLA achieves 'pure on-body edge inference without relying on cloud computing,' but training and group updates still rely on the cloud. Zhongqing researchers say they aim to build a unified base model, providing personalized capabilities for different products in the future.

Zhao Tongyang said that now they can finally calculate the ROI of humanoid robots. In some sorting scenarios, robot efficiency and accuracy approach human levels, and some projects can already calculate investment returns. On the supply side, costs are dropping rapidly: ball screw prices have fallen from 10,000-20,000 yuan three years ago to 1,000-2,000 yuan, and reducer prices have generally halved.

Morgan Stanley has raised its forecast for China's humanoid robot shipments: from 28,000 units in 2026 to 50,000 units, with 446,000 units expected by 2030, corresponding to a 106% compound annual growth rate from 2025 to 2030. But Zhao Tongyang believes the industry has not yet exploded; only when a company produces hundreds of thousands or millions of units annually can it be considered a boom, and the current scale is still in the tens of thousands.

Zhongqing invests 80%-90% of resources in commercial, industrial, and home scenarios, with combat competitions accounting for only 10%. In August, the T800 entered Luxshare Precision's Suzhou factory. The Shenzhen base covers about 12,000 square meters, with a takt time of one unit every 15 minutes and 79 quality checks; in July, the first batch of T800 units rolled off the Zhengzhou base. Zhao Tongyang said the cost of general-purpose humanoid robots for deployable scenarios has been controlled to the hundred-thousand-yuan level, but due to lack of mass production, there is still room for further reduction.

The market Zhao Tongyang truly wants to win is the home, a judgment he made when he entered the robotics industry 10 years ago. He failed twice in entrepreneurship due to technical difficulty and funding pressure. After founding Dogotix in 2019, it was acquired by He Xiaopeng. In 2023, he left due to disagreements over the direction, taking only about 1.2 million yuan, and founded Zhongqing in October of the same year.

Zhao Tongyang describes himself as a 'person with three lacks': no money, no halo, no management skills. His entrepreneurial journey is similar to a combat league: repeatedly getting punched and standing up again. These lessons are written into Zhongqing: products must be eye-catching and cheap, and release, financing, supply chain, and mass production must all be fast.

This year, Zhao Tongyang is trying to use 'ecosystem' to handle complexity, like Nvidia, building a good underlying platform and leaving specific tasks to users and partners. But the premise is that shipment volumes and developer activity can support it; otherwise, the ecosystem is just a list of partners. Zhao Tongyang said this year's humanoid robot delivery volume is expected to be four to five times last year's, with deliveries aimed at high value, not just to boost revenue.